OpenAI Security Warnings: What Reuters Reported and What Changed

On Oct. 13, the Justice Department said federal agents had arrested a former OpenAI engineer, accused of hiding an audio recorder, concealed in toilet paper, in an executive conference room. According to Reuters, several former employees had the same reaction when OpenAI disclosed the device: they weren't surprised, and Reuters now reports that OpenAI security warnings from staff go back years.

That is the starting point for the OpenAI security warnings in a Reuters investigation by Deepa Seetharaman, published Nov. 3. OpenAI, the company behind ChatGPT, is the subject, and its security practices are the story. We read the report in a syndicated copy on Claims Journal; Reuters' own page didn't load for us.

Who told Reuters what

The report rests on anonymous sources, and the number behind each claim matters. Five people familiar with the matter told Reuters that staff raised security concerns internally even before ChatGPT launched in late 2022, took them to top leaders and the board, and that some left those conversations believing management cared more about keeping pace with Google and Anthropic than about security.

Three people said OpenAI hadn't done regular security reviews of internal tools such as the sandboxes where researchers stress-test new models. Two said Thibault Sottiaux, a senior executive overseeing product development, recently took charge of an effort to secure internal systems. One said that in August he moved a quarter of OpenAI's production engineers onto security infrastructure. That last figure has the thinnest sourcing in the piece, and OpenAI didn't confirm it in the quoted response.

OpenAI's spokesperson, Elana Widmann, said the company is "one of the most targeted organizations in the world" and called protecting its technology "a matter of national importance." She said recent actions show the security program "doing its job: identifying risks and responding decisively."

The monorepo, and why agents change the math

The technical complaint is about access. Reuters says OpenAI kept most of its code in a single "monorepo," so anyone who could reach it could see the details of many projects, including sensitive ones unrelated to their work. The report adds a detail that explains the timing: the repository was dense and complicated, which may have deterred some would-be attackers in the past, but AI agents can now sift that much material.

OpenAI pushed back on the scale. It said much of its most sensitive work sits in separate repositories with strict access controls, and that the monorepo doesn't contain all of its code or the model weights, the numerical values that determine how a model behaves.

Reuters also lists earlier incidents. After a 2023 outside hack that took details about GPT-4, OpenAI said the intruder reached only Slack channels open to the whole company. This summer, OpenAI agents being tested hacked into Hugging Face. And in July, researchers at Hacktron AI used an Anthropic model to get into ChatGPT accounts belonging to several OpenAI employees and reached some internal systems, including the monorepo, according to the security firm, as Reuters reports it. OpenAI said it thanked the researchers and took steps to reduce the risk of similar attacks.

OpenAI's response, in numbers

After Reuters asked for comment, OpenAI posted on X late on Nov. 2 US time (02:29 UTC on Nov. 3) about its "next internal security sprint," run with a partner, Cathedral Systems, "whose researchers will use our models alongside our teams to challenge our defenses." The post also says rewards for critical findings are more than quadrupling, with exceptional findings eligible for up to $172,000, including bonuses.

OpenAI on its security sprint and bug bounty, Nov. 3, 2026 (UTC) (@OpenAI)

View on X

The post gives no start date, duration or list of systems in scope. The $172,000 is a ceiling for exceptional findings, not a payout.

One name appears twice

Cathedral Systems shows up in the Reuters report for a different reason. In late October, it says, the cybersecurity startup claimed it had hacked into Anthropic's systems and reached Slack channels where employees discussed sensitive model details. An Anthropic spokesperson told Reuters the data accessed included some business contact information for customers and partners and a set of basic user account records.

Neither Reuters nor OpenAI's post connects that episode to the sprint, and we don't know the terms of the partnership. It's a fact worth keeping in view: the outside team invited to test OpenAI's defenses had just publicized a breach of a rival lab.

The quote behind the headline

Reuters cites a post by Will DePue, a researcher who left OpenAI in April, arguing that AI leaders "can't sacrifice velocity for security." It's an opinion about industry competition, not a finding about OpenAI's systems.

Will DePue's post on X, Oct. 28, 2026 (UTC) (@willdepue)

View on X

If your own agents read your repo

Nothing in these documents changes OpenAI's APIs, pricing or model availability. The part that reaches other teams is the observation about monorepos: if a coding agent can index everything you can, a large repository is no longer protected by being hard to read. Check what your agents can see before you give them the whole tree, and keep secrets, credentials and internal tooling in separate repositories with their own permissions.

OpenAI's spokesperson said the program is working. The measure will be what the sprint finds and whether OpenAI says so, and the company hasn't yet given a date to check.

Source: Reuters

openai openai security reuters security sprint cathedral systems

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.